Omnidirectional, 360 \(^\circ \) or spherical images are signals defined on the sphere surface that capture the entire field of view. High-resolution 360 \(^\circ \) images are applicable in virtual reality and require efficient compression. Spherical images are commonly mapped to the plane using the equirectangular projection, which suffers from latitude-dependent distortions. Existing 360 \(^\circ \) image block-based compression methods introduce blocking artifacts, degrading image quality, or face design issues that might turn their usage impractical. This paper proposes a new content-guided compression method designed for spherical images that uses spherical superpixels as coding units. The proposed method applies a spherically-weighted graph Fourier transform and a latitude-adaptive quantization scheme to each superpixel. The proposed approach enables state-of-the-art compression ratios with results up to 3.66dB in image quality and -55% of bit rate regarding other recent coding unit-based 360 \(^\circ \) image compression techniques.
